Partial Least Squares Regression on Grassmannian Manifold for Emotion Recognition

Transcription

1 Emotion Recognition In The Wild Challenge and Workshop (EmotiW 2013) Partial Least Squares Regression on Grassmannian Manifold for Emotion Recognition Mengyi Liu, Ruiping Wang, Zhiwu Huang, Shiguang Shan, Xilin Chen Institute of Computing Technology, Chinese Academy of Sciences

2 Outline Problem Related work Our Method Experiments Conclusion 2

3 Outline Problem Related work Our Method Experiments Conclusion 3

4 Emotion recognition in the wild Challenges Large data variations head pose, illumination, partial occlusion, etc. Lack of labeled data Manual annotation is hard as spontaneous expression is ambiguous in the real world. 4

5 Outline Problem Related work Our Method Experiments Conclusion 5

6 Video-based emotion recognition Acoustic information based Time domain and frequency domain e.g. pitch, intensity, pitch contour, Low Short-time Energy Ratio (LSTER), maximum bandwidth, Vision information based Spatial space and temporal space e.g. Optical flow, 3D descriptor (LBP-TOP, HOG 3D), tracking based (AAM, CLM), probabilistic graph model (HMM, CRF), 6

7 Outline Problem Related work Our Method Experiments Conclusion 7

8 Key issue Our method How to model the emotion video clip? Motivation Alleviate the effect of mis-alignment of facial images Encode the data variations among video frames Basic idea Inspired by recent progress of image set-based face recognition [1] Treat the video clip as an image set, i.e., a collection of frames Linear subspace for video (image set) modeling 8 [1] R. Wang, H. Guo, L. S. Davis, and Q. Dai. Covariance discriminative learning: A natural and efficient approach to image set classification. CVPR, 2012.

9 Our method An overview Preprocessing Feature Designing Classification Original aligned face images Purified face images Mid-level image features Video/Image set features One-to-Rest PLS classification Video Filtering out non-face in PCA subspace Subspace learning on Grassmannian manifold Video-Audio Fusion Audio Original audio data Clip-wise audio features extracted using opensmile toolkit*[2] One-to-Rest PLS classification 9 [2] F. Eyben, M. Wollmer, and B. Schuller. Opensmile: the munich versatile and fast open-source audio feature extractor. ACM MM, 2010.

10 Preprocessing Our method Original face alignment using MoPS [3] (provided by organizer) Purification of face images Original aligned face images set: X = x 1, x 2,, x n, x i R D. PCA projection learned on X by preserving low energy: W. Mean reconstruction error of each image: MeanErr t = 1 D x t W T 2 Wx t Non-face/Badly-aligned face images tend to have large MeanErr t. 10 [3] X. Zhu, and D. Ramanan. Face detection, pose estimation, and landmark localization in the wild. CVPR, 2012.

11 The Number of Samples Preprocessing Our method The distribution of MeanErr t on training set in EmotiW Threshold The Mean Reconstruction Error * Threshold is for filtering out non-face in PCA space. 11

12 Our method Preprocessing An example of 100 samples with largest mean reconstruction error. Most are non-face images or mis-alignment results. 12

13 Our method An overview Preprocessing Feature Designing Classification Original aligned face images Purified face images Mid-level image features Video/Image set features One-to-Rest PLS classification Video Filtering out non-face in PCA subspace Subspace learning on Grassmannian manifold Video-Audio Fusion Audio Original audio data Clip-wise audio features extracted using opensmile toolkit*[2] One-to-Rest PLS classification 13

14 Feature designing Image feature [4] Our method Convolution Filters 6x6x100 Max-Pooling 3x3 Face Image 32x32 Filter Maps 27x27x100 Mid-level Feature 9x9x100 [4] M. Liu, S. Li, S. Shan, X. Chen. AU-aware Deep Networks for Facial Expression Recognition. FG,

15 Our method Feature designing Video feature Each video clip is a set of images, denoted as S i R f n i, where f is the dimension of image feature, and n i is the number of frames. The video S i can be represented as a linear subspace P i, s.t. S i S i T = P i Λ i P i T Thus all the video clips can be modeled as a collection of subspaces, which are also the points on Grassmannian manifold. 15

16 Feature designing Video feature Our method An illustration of subspaces on Grassmannian manifold Video Clip 1 Video Clip 2 Similarity M 16

17 Our method Feature designing Video feature The similarity between two points P i and P j on manifold M can be measured by a linear combination of Grassmannian kernels. Projection kernel[5]: k proj ij = P T 2 i P j F. Canonical correlation kernel[6]: k CC ij = max ap span P i max bq span P j a T p b q. Linear combination: k com ij = k proj ij + αk CC ij. The kernels of each point (i.e., each video) to all training points serve as its final feature representation for classification. [5] J. Hamm, D. Lee. Grassmann discriminant analysis: a unifying view on subspace-based learning. ICML, [6] M. Harandi, C. Sanderson, S. Shirazi, B.C. Lovell. Graph embedding discriminant analysis on Grassmannian manifolds for improved image set matching. CVPR,

18 Our method An overview Preprocessing Feature Designing Classification Original aligned face images Purified face images Mid-level image features Video/Image set features One-to-Rest PLS classification Video Filtering out non-face in PCA subspace Subspace learning on Grassmannian manifold Video-Audio Fusion Audio Original audio data Clip-wise audio features extracted using opensmile toolkit*[2] One-to-Rest PLS classification 18

19 Our method Classification Partial Least Squares (PLS) for classification [1] Maximize the covariance between observations and class labels Space X Space Y Each sample is a feature vector Each sample is a 0/1 class label Linear Projection Latent Space T X = T * P Y = T * B * C = X * B_pls T is the common latent representation [1] R. Wang, H. Guo, L. S. Davis, and Q. Dai. Covariance discriminative learning: A natural and efficient approach to image set classification. CVPR,

20 Classification 20 One-to-Rest PLS Our method Suppose there are c categories and N training samples, we train c One-to-Rest PLS classifiers to predict each class simultaneously. Effectively to handle the hard classes, e.g. Sad vs. Disgust Binarize Original training label vector Y R N 1 Binary training label matrix Y R N c Separate One-to-Rest training label vectors, y 1, y 2,, y c R N 1

21 Our method Classification One-to-Rest PLS Training and test process Training data X + One-to-Rest training label vectors y 1, y 2,, y c Test sample One-to-Rest PLS(1) One-to-Rest PLS(2) One-to-Rest PLS(3) One-to-Rest PLS(c-1) One-to-Rest PLS(c) Test result: Fit R c 1 21

22 Our method Classification Video-Audio fusion for final test output For a given test video, using the c PLS classifiers for video and audio respectively, we obtain two prediction vectors Fit video, Fit audio R c 1. We conduct a linear fusion at decision level using weighted parameter λ Fit fusion = (1 λ) Fit video +λfit audio. The category corresponding to the maximum value in Fit fusion is determined to be the recognition result. 22

23 Outline Problem Related work Our Method Experiments Conclusion 23

24 Recognition Accuracy Experiments Discussion of Parameters The fusion weights of Grassmannian kernels Train-Val Val-Train k ij com = k ij proj + αk ij CC = 2 6, 2 5 = The Value of Grassmannian Kernels Combination Weight (Alpha) 24

25 Recognition Accuracy (Video) Experiments Discussion of Parameters The dimension of One-to-Rest PLS (video) Train-Train Train-Val Val-Val Val-Train = 10 = The Dimensions of PLS 25

26 Recognition Accuracy (Audio) Experiments Discussion of Parameters The dimension of One-to-Rest PLS (audio) Train-Train Train-Val Val-Val Val-Train = 5 = The Dimensions of PLS 26

27 Recognition Accuracy Experiments Discussion of Parameters The fusion weights of video and audio modalities Train-Val Val-Train Fit fusion = (1 λ) Fit video +λfit audio = = The Value of Video-Audio Fusion Weight (Lambda) 27

28 Experiments Results comparison Performance Comparison Audio only One-to-Rest PLS Grassmannian Discriminant Analysis [6] Video only Grassmannian Kernels + One-to-Rest PLS Original data Feature-level fusion Multi-class LR One-to-Rest PLS Audio + Video Decisionlevel fusion One-to-Rest PLS Purified data Decisionlevel fusion One-to-Rest PLS Ours Baseline Val % 30.81% 32.07% 22.48% 24.24% 34.34% 35.86% Test* % % 33.01% 34.61% Val 19.95% 27.27% 22.22% Test 22.44% 22.75% 27.56% [6] M. Harandi, C. Sanderson, S. Shirazi, B.C. Lovell. Graph embedding discriminant analysis on Grassmannian manifolds for improved image set matching. CVPR,

29 Outline Problem Related work Our Method Experiments Conclusion 29

30 Conclusion Key points of the current method PCA-based data purifying to filter out mis-alignment faces Linear subspace modeling of video data variations Multiple video features fusion by Grassmannian kernels combination Multi-modality fusion at decision level of video and audio Issues to further address 30 Exploration of video temporal dynamics information More sophisticated video modeling More effective fusion at feature level

31 Thank you. Question?